SF Express

Qingpu District, Shanghai, China

Customized YOLOv8 for object detection and multi-object tracking, running controlled experiments across attention modules, detection heads, and training configurations, then optimized the resulting models for deployment.

• Customized YOLOv8 with PyTorch/Ultralytics by integrating SE, CBAM, ECA, PSA, DCN, and dynamic convolution into C2f architectures; achieved +8 pp mAP with CBAM on a custom vision dataset.
• Evaluated detection heads, label assignment strategies, loss functions, optimizers, and LR schedulers using controlled experiments, identifying performance and convergence trade-offs across model configurations.
• Integrated ByteTrack for multi-object tracking and analyzed accuracy–latency trade-offs; identified an 18% inference-time overhead from DCN-based architectures.
• Applied structured/unstructured pruning and FP16 quantization, achieving 21% lower GPU memory usage and 59% higher inference throughput with negligible accuracy loss.
• Extended YOLOv8 to instance segmentation using YOLOv8n-seg, developed data augmentation pipelines with Albumentations and PyTorch DataLoader, evaluating flipping, rotation, scaling, and other augmentation strategies to improve model generalization.

UTMIST AI Squad

Toronto, Ontario

Built a foundation in machine learning through model training, evaluation, and hands-on projects.

• Built a foundation in machine learning through model training, evaluation, and practical hands-on projects.
• Collaborated with teammates and independently researched technical documentation to solve implementation challenges.

UTMIST Eigen AI

Toronto, Ontario

Presented Zero-Trust security concepts and modern network security principles to a non-technical audience.

• Presented Zero-Trust security concepts and modern network security principles to a non-technical audience.